Chris builds a calorie tracking app in 12 hours to experiment with Apple’s liquid glass design and foundation AI models, discovering that while local AI models offer smooth performance, they lack accuracy in calorie estimation. By integrating cloud-based AI and nutrition databases, he achieves better results, ultimately creating a promising app that garnered social media interest and plans for a beta release.
In this video, Chris, a developer known for building productivity apps, takes a break from his main project, Subscription Monster, due to burnout and decides to build a fun, experimental app in 12 hours. His goal is to explore two new Apple technologies: the recently released liquid glass design and Apple’s foundation models, which are AI models that run locally on devices. He chooses to create a calorie tracking app with a unique twist—users simply type the food they eat, and the AI estimates the calories without relying on databases or barcode scanning.
Chris begins by designing the app’s user interface using the liquid glass aesthetic, which he finds visually appealing and plans to incorporate into his other projects. He demonstrates the app’s basic functionality, including a settings page and a daily calorie goal feature. The app uses Apple’s foundation models to estimate calories based on user input, but Chris quickly discovers that while the models run smoothly on the device, their calorie estimations are often inaccurate, such as overestimating the calories in an In-N-Out Burger.
To improve accuracy, Chris experiments with other local AI models from HuggingFace, including Meta’s Llama and Google’s GMA models, using a Swift package to integrate them. Although these models offer slightly better accuracy, they come with significant performance drawbacks compared to Apple’s optimized foundation models. He then tests cloud-based AI providers like OpenAI and Anthropic, which show improved accuracy but still struggle with specific food items. To address this, Chris integrates the Fat Secret nutrition database API, combining it with AI to extract food items and portions and fetch precise nutrition data, resulting in much better calorie estimates.
However, the Fat Secret API is costly, with commercial use starting at $1,500 per month, and requires branding on the app, which Chris wants to avoid for aesthetic reasons. Continuing his research, he discovers the Perplexity Sonar model, which combines AI with real-time internet search capabilities. This model provides highly accurate calorie information by pulling data from multiple sources, including restaurant websites, and even shows the sources used, increasing user trust. Despite its higher cost per request, Chris considers it the best solution for his app’s needs.
In conclusion, Chris reflects on the project as a valuable learning experience, especially in understanding Apple’s foundation models and liquid glass design. Although the foundation models were not ideal for calorie estimation, the app turned out better than expected and generated significant interest on social media, prompting him to open a waitlist for potential users. He plans to release a beta version soon and expresses enthusiasm for continuing to explore innovative app ideas, encouraging viewers to follow his journey on social media and subscribe for more content.